Of the 14,130 replies we read, 922 sit under a review that is not in English. Arabic, Norwegian, German, Spanish, Italian, Polish. Someone came in, liked it or did not, and wrote about it in their own language.
788 of those 922 — 85 per cent — were answered in English.
Why this is not a small thing
A reply has two audiences. The reviewer, who may or may not come back, and every future customer scrolling the listing. Answering in English still serves the second audience — but it tells the first one, plainly, that the message did not land.
In a London café with a hotel two streets away, or a Manchester café near a university, the share of reviews in another language is not a rounding error. It is a recurring slice of the people paying you.
What most of these replies actually say
Almost all of them are the generic thank-you. That is not a coincidence: faced with a review they cannot read, the owner falls back on the safest sentence they have. Which means the language problem and the generic-reply problem are the same problem wearing different clothes.
“Thank you for your review! We hope to see you again soon.”
The measurement problem this creates for us
We measure whether a reply engages with the review by looking at word overlap. Across two languages that test returns zero every time — not because the reply is bad, but because the test cannot see. We found this the hard way: a reply that correctly picked up three separate points from a German review scored nought.
So our own tools now say so. Paste a non-English review into our checker and it tells you it cannot measure this honestly, rather than printing a number it cannot stand behind. A measurement that reports its own blind spot is worth more than one that does not have any.
What to do about it
Reply in the reviewer's language, and keep the English. Two short paragraphs, the reviewer's language first. The reviewer reads the first; every future customer reads the second; the listing shows a business that noticed.
It costs nothing and almost nobody does it. Of the reviews in our corpus written in another language, the number answered in that language is small enough that a café doing it consistently would stand out on its own street.
Where these reviews come from
922 of 14,130 is about one review in fourteen. That is not evenly spread: a café by a station or a university sees far more of them than one on a residential street, and for those businesses this is not an edge case but a weekly occurrence.
The scripts we found most often are the ones you would expect from a country that hosts fifty million visitors a year — and the replies underneath them are, almost without exception, the same English thank-you.
The cost of the safe sentence
This connects to the larger finding in our corpus: 51 per cent of all replies share no distinctive word with the review above them. A reply written to a review the owner could not read is guaranteed to fall in that group. The language barrier does not create the generic reply — it just removes the last reason not to write one.
What it would take
Very little. Two short paragraphs, the reviewer's language first, English second. No translation service, no new software — the reply box takes any characters you paste into it.
The reason to bother is not politeness. It is that the listing is read by everyone who comes after, and a business that answered someone in Italian is visibly a business that read the review.
The wider pattern this belongs to
Answering in the wrong language is one version of a bigger habit. Across the whole corpus, 3,018 replies — 21 per cent — are eight words or fewer, and 664 sentences appear word for word at two or more separate businesses. The reflex in all three cases is the same: reach for the sentence that cannot cause trouble.
It is a rational reflex. The cost of a careless reply is visible and the cost of an empty one is not. But the empty one is paid every time somebody scrolls past.
A note on what we can and cannot say here
We can say how many reviews were not in English and how many were answered in English. We cannot say what those reviewers thought about it, whether they came back, or whether a reply in their own language would have changed anything. Nobody has that data, and we are not going to model it.
What we can say is that the reply is public and permanent, and that a business which answered someone in Italian is visibly a business that read the review.
One more thing the timestamps show
Reviews do not arrive during office hours. Of the 12,022 reviews in our corpus carrying a usable timestamp, 6 per cent were written between ten at night and six in the morning. The person writing them is at home, on a phone, after the visit — and often after a few hours' distance from it.
That is worth holding alongside the language finding. The reply is not part of a conversation happening in the café. It is written hours or days later, to someone who has already left, and read mostly by people who were never there.
How this was measured
14,130 review-and-reply pairs from 914 UK cafés, August 2026. A review counts as not-English when it uses a non-Latin script, a Scandinavian character, or a function word that does not occur in English. Loanwords that English has absorbed — café, crème, béarnaise — are stripped before the test, because an English review that mentions a béarnaise is still an English review.